Can a brain-to-image reconstruction look correct even when the brain signal contains no meaningful information?
This repository introduces Brain Control (BC), a metric designed to quantify how much reconstructed images are actually driven by neural activity rather than by powerful generative priors.
Our key finding is striking:
A modern reconstruction pipeline can generate realistic, category-consistent images even when the underlying brain signals are completely randomized—and standard evaluation metrics often fail to detect it.
Recent brain decoding systems can reconstruct visually impressive images from fMRI data using diffusion models, GANs, and large-scale visual priors.
However, visual realism does not necessarily imply neural control.
A reconstruction can appear successful because the generator already knows how to produce plausible natural images, regardless of whether the decoded brain activity contains stimulus-specific information.
This creates a fundamental evaluation problem:
| Condition | BC | Feature Norm | Visual Quality |
|---|---|---|---|
| Real fMRI | 1.259 | 70.9 | Plausible |
| Shuffled fMRI | 1.001 | 70.9 | Plausible |
| Gaussian Noise | 1.000 | 118.4 | Plausible |
Even after destroying all brain–stimulus correspondence by shuffling fMRI signals:
- reconstructed images remain visually plausible,
- decoded feature magnitudes remain unchanged,
- standard metrics show little indication of failure.
Brain Control (BC) detects this failure mode directly.
BC measures how much structure disappears when the relationship between brain activity and stimulus labels is intentionally broken.
[ BC = \frac{V_{\text{brok}}}{V_{\text{pres}}} ]
where:
- (V_{\text{pres}}): within-category feature variance under preserved correspondence
- (V_{\text{brok}}): expected variance after random shuffling of brain–stimulus mappings
-
BC = 1
- Shuffling has no effect
- Reconstruction is prior-dominated
- Brain signals contribute little or no information
-
BC > 1
- Preserved correspondence reduces variance
- Brain activity imposes meaningful structure
- Reconstruction is genuinely brain-driven
Under a Gaussian approximation:
[ I(c;\hat{x}) \approx \frac{1}{2}\log(BC) ]
Thus:
- BC = 1 implies approximately zero mutual information between category labels and decoded representations.
- Larger BC values indicate stronger neural control.
Real fMRI:
- BC = 1.259
Shuffled fMRI:
- BC = 1.001
Yet decoded feature norms are identical:
- Real: 70.9
- Shuffled: 70.9
This demonstrates a critical dissociation:
standard reconstruction statistics remain unchanged while neural control disappears.
| Subject | Real BC | Shuffled BC |
|---|---|---|
| 1 | 1.259 | 1.001 |
| 2 | 1.135 | 0.999 |
| 3 | 1.250 | 0.999 |
| 4 | 1.208 | 1.001 |
| 5 | 1.138 | 1.002 |
Paired t-test:
- t = 7.44
- p = 0.0017
Across ROIs:
- Pearson r = 0.967 between BC and identification accuracy
However:
- matched-accuracy comparisons still show significant BC differences
- Cohen's d = 0.465
- p = 0.023
This indicates BC is not simply another accuracy metric.
Instead, it measures a distinct property:
how strongly brain activity constrains the reconstruction process.
| Figure | Description |
|---|---|
| fig1_concept.png | Conceptual illustration of prior dominance |
| fig2_reconstruction_comparison.png | Real vs Shuffled vs Random reconstructions |
| fig3_bc_barplot.png | BC and feature norm comparison |
| fig4_noise_sensitivity.png | Noise sensitivity analysis |
| fig5_bc_vs_accuracy.png | BC vs identification accuracy |
pip install numpy scipy scikit-learn matplotlib pillow h5pypython experiments/exp13_prior_vs_brain.pypython experiments/exp17_multisubject_bc.pypython experiments/exp15_roi_bc_vs_accuracy.pyfrom compute_bc import compute_bc
bc_mean, bc_sem, bc_per_cat = compute_bc(
pred_features,
cat_labels,
n_shuffle=1000,
seed=42,
mode="across"
)
print(f"BC = {bc_mean:.4f} ± {bc_sem:.4f}")Interpretation:
BC ≈ 1.0 → Prior-dominated reconstruction
BC > 1.0 → Brain-driven reconstruction
bibtex @article{kikuchi2026bc, title={Plausible but Not Brain-Driven: Quantifying Prior Dominance in Neural Image Reconstruction}, author={Kikuchi, Yuki}, year={2026}, note={Preprint} }
(arXiv link coming soon)
MIT License